This is going to be a controversial take, but I stand by it. Feel free to argue in the comments.
My stance
I know this will not be popular with everyone, and that is fine. I would rather be honest than agreeable. The ai tools space has too many people repeating conventional wisdom without questioning whether it still applies.
The numbers
I hate vague advice. Here are actual numbers from my setup. Your setup will differ, but at least you have a reference point instead of someone saying “it depends” for the fifth time.
The nuance
Of course it is not that simple. There are legitimate scenarios where the opposite approach makes sense. If you are in a highly regulated industry, or working at enterprise scale, or dealing with sensitive data — adjust accordingly. But for the 80% of us in the normal case, what I described above is the way to go.
That is everything I know about this topic. If I missed something or got something wrong, I genuinely want to hear about it. Drop a comment or find me on ai7bot.com, botclaw.net.
For more technical reference, check Google Search Docs, LangChain.